Automatic detection of optic canal fractures and recognition and segmentation of anatomical structures in the orbital apex based on artificial intelligence.

Abstract

Background and objectives: Traumatic optic neuropathy (TON) caused by optic canal fractures (OCF) can result in severe visual impairment, even blindness. Timely and accurate diagnosis and treatment are crucial for preserving visual function. However, diagnosing OCF can be challenging for inexperienced clinicians due to atypical OCF changes in imaging studies and variability in optic canal anatomy. This study aimed to develop an artificial intelligence (AI) image recognition system for OCF to assist in diagnosing OCF and segmenting important anatomical structures in the orbital apex.

Methods: Using the YOLOv7 neural network, we implemented OCF localization and assessment in CT images. To achieve more accurate segmentation of key anatomical structures, such as the internal carotid artery, cavernous sinus, and optic canal, we introduced Selective Kernel Convolution and Transformer encoder modules into the original UNet structure.

Results: The YOLOv7 model achieved an overall precision of 79.5%, recall of 74.3%, F1 score of 76.8%, and mAP@0.5 of 80.2% in OCF detection. For segmentation tasks, the improved UNet model achieved a mean Intersection over Union (mIoU) of 92.76% and a mean Dice coefficient (mDice) of 90.19%, significantly outperforming the original UNet. Assisted by AI, ophthalmology residents improved their diagnostic AUC-ROC from 0.576 to 0.795 and significantly reduced diagnostic time.

Conclusion: This study developed an AI-based system for the diagnosis and treatment of optic canal fractures. The system not only enhanced diagnostic accuracy and reduced surgical collateral damage but also laid a solid foundation for the continuous development of future intelligent surgical robots and advanced smart healthcare systems.